AI-ordlista
Den kompletta ordlistan över AI
Hybrid Diffusion-GAN
Architecture combining diffusion models and generative adversarial networks to benefit from the training stability of GANs and the generative quality of diffusion models. This hybrid approach enables high-resolution image generation with fine control over the generation process.
Diffusive GAN
A variant of GANs where the generator uses a progressive diffusion process to transform noise into realistic data, while the discriminator evaluates intermediate states. This method improves the diversity of generated samples and reduces mode collapse.
Adversarial Diffusion
A diffusion process where denoising is guided by an adversarial mechanism, using a discriminator to assess realism at each step of the reverse process. This approach combines the robustness of diffusion models with the finesse of GANs.
Diffusion-Conditioned GAN
An architecture where a diffusion model conditions the GAN, providing a structured latent representation that guides the generation of both the discriminator and generator. This method allows precise control over generated attributes while maintaining overall coherence.
Hybrid Score Matching
A technique combining score estimation from diffusion models with the adversarial loss function of GANs to jointly optimize generative quality and diversity. This hybrid approach improves training convergence and stability.
Adversarial Diffusion Process
A methodology where the noise addition and denoising process is supervised by a discriminator that assesses quality at each timestep of the forward and backward process. This continuous supervision ensures more realistic transitions between latent states.
Diffusion-Step GAN
An architecture where the GAN generator operates on successive steps similar to the timesteps of a diffusion model, progressively refining the output. This approach allows granular control over generation quality at each step.
Diffusion Model with Discriminator
A variant of diffusion models incorporating a discriminator in the denoising process to evaluate and improve the realism of intermediate samples. This hybridization maintains the advantages of diffusion while adding an adversarial quality criterion.
Structured Noise GAN
GAN architecture using structured initial noise generated by a diffusion process, rather than pure random Gaussian noise. This approach provides the generator with a more consistent starting point for generating complex images.
Two-Phase Diffusion
Hybrid model where a first diffusion phase creates a basic latent representation, followed by a GAN phase that refines and adds high-frequency details. This division of labor optimizes computational efficiency and final quality.
Continuous-Time GAN
Extension of GANs where time is treated as a continuous variable, similar to diffusion models, allowing smooth interpolation between different generation states. This approach facilitates temporal control and generation consistency.
Stochastic Adversarial Diffusion
Method combining the stochastic nature of diffusion models with adversarial optimization, where each denoising step is evaluated by a probabilistic discriminator. This approach maintains diversity while ensuring quality.
Diffusive Latent Space GAN
Architecture where the GAN's latent space is structured according to diffusion model principles, with a hierarchy of controlled noise levels. This organization enables better interpolation and more intuitive generation control.
Hybrid DDPM-GAN Model
Specific combination of Denoising Diffusion Probabilistic Models with GAN architectures, using DDPM for base generation and GAN for final refinement. This approach leverages DDPM's stability and GAN's sharpness.
Diffusion with Adversarial Loss
Modification of diffusion models where the standard loss function is complemented by an adversarial loss, evaluated by a simultaneously trained discriminator. This hybridization improves the sharpness and realism of final generations.